Research Information System
9th International Conference on Spoken Language Processing/INTERSPEECH 2006, Pennsylvania, United States Of America, 01 January 2006, pp.2378-2380
Naturalness of synthetic speech highly depends on appropriate modeling of prosodic aspects. Mostly, three prosody components are modeled: segmental duration, pitch contour and intensity. In this study, we present our work on modeling segmental duration in Turkish using machine-learning algorithms, especially Classification and Regression Trees (CART). The models predict phone durations based on attributes such as phone identity, neighboring phone identities, lexical stress, position of syllable in word, part-of-speech (POS) information, word length in number of syllables and position of word in utterance extracted from a speech corpus of approximately 700 sentences. Obtained models predict segment durations better than mean duration approximations (similar to 0.77 Correlation Coefficient, CC, and 20.4 ms Root-Mean Squared Error, RMSE). Attributes phone identity, neighboring phone identities, lexical stress, syllable type, POS, phrase break information, and location of word in the phrase constitute best predictor set for phoneme duration modeling.